Trans-CNN: A Transfer Learning Based Approach for Dolphin Sound Recognition

Due to the uneven characteristics of the data collected by Passive Acoustic Monitoring (PAM) and the lack of effective labels, it has been difficult to accurately identify the acoustic signals of marine mammals. Aiming at these well-known challenges, in this study, we integrate convolutional neural networks and, with the help of transfer learning strategies, construct a model specifically for dolphin sound signal recognition. The proposed network is called transfer learning-based convolutional neural network (Trans-CNN). We first construct a Pre-train Dataset containing labels based on the audio signals in the Watkins database, and an unlabeled dataset (Dolphin Dataset) for transfer learning. Next, these datasets are applied on the Trans-CNN model. The experimental results show that for the recognition tasks of 12 different dolphin sounds, Trans-CNN has achieved an average accuracy of 90%, which is more than 15% higher than the traditional recognition method.